Next-Generation Peptides: AI-driven approaches for peptide therapeutics beyond the natural repertoire

Doctoral thesis, 2026

Peptides are becoming an attractive modality in drug discovery as they sit in the intersection of small molecules and proteins. They combine advantages drawn from both modalities such as high specificity, low immunogenicity typically associated with protein therapeutics, and good efficacy, potential for membrane permeability more common to small molecules. Nevertheless, they demand an exhaustive search for an optimal amino acid sequence to meet a multi-objective profile including metabolic stability, solubility, and potency needed for becoming a standalone drug. Essential aspect of property optimization is incorporating non-natural amino acids (NNAAs) to peptides, generally to enhance their permeability and affinity. However, design make-test-analyze (DMTA) cycles often rely on trial‑and‑error of positional mutations. This makes the identification of peptides meeting the design goals exhaustive and time‑consuming. Drug discovery pipelines are accelerated by the recent surge of artificial intelligence (AI)-driven technologies for small molecules and proteins. This thesis presents in silico tools that facilitate drug design by extending AI-based methodologies to peptide therapeutics. The solutions presented here enable designs beyond the natural amino acids while allowing efficient exploration of the chemical space that is expanded by novel and diverse NNAAs. This includes developing AI-driven methodologies for design, evaluation, and optimization of next-generation therapeutic peptides. To design peptides, a chemistry-aware generative model was built to incorporate NNAAs into user-defined positions of a given starting peptide. This model is guided by reinforcement learning feedback to iteratively optimize designs for desired properties such as permeability and solubility. This design process is supplemented by a series of methodologies to evaluate the generated designs. First, peptide-specific predictive models that leverage model uncertainty were developed to efficiently predict permeability and steer design decisions toward reliable property space. Second, an NNAA synthesis assistance tool was proposed. This tool evaluates the chemical synthesizability of amino acids by considering protection strategies required for peptide synthesis and adapts predictive models for small molecule retrosynthesis and synthetic feasibility to peptide building blocks. Collectively, studies presented in this thesis develop cheminformatics and AI applications to design novel, synthesizable and pharmacologically relevant peptides while expanding the chemical space accessible to peptide drug discovery.